Skip to content

Release 2.90.0 - #1101

Merged
adibhanna merged 8 commits into
mainfrom
v2.90.0
Nov 20, 2026
Merged

adibhanna merged 8 commits into
mainfrom
v2.90.0

Conversation

@adibhanna

@adibhanna adibhanna commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor

ZenNotes 2.90.0: on Windows, a note made again under a deleted note's name keeps its own date; zn 0.18.1; KaTeX 0.19.1; Windows CI and test fixes.

The renumber tests failed now and then on slow runners, last on Windows
x64 in PR #1098: "pads every marker once the list outgrows the column
(99 to 100)" kept the old padding. makeState forced a full parse with
ensureSyntaxTree, but that finishes the parse inside the language
state's context and leaves the tree syntaxTree(state) returns at the
one the state was created with. lang-markdown's
insertNewlineContinueMarkup reads syntaxTree(state), so when the
creation parse ran out of time partway down a long list, Enter did not
see the list at all. On a 200,000-line document the creation tree
covered 3,002 characters and still did after ensureSyntaxTree; an empty
transaction made it cover all 3,777,784.

makeState now returns state.update({}).state after the forced parse.
The app is not affected: a view keeps applying the parse worker's tree.
Windows x64 failed main at abf115d and three runs of PR #1098 on
desktop tests that hit vitest's 5 s limit: the MCP "update_task on a
recurring task" cases, the vault demo tour, asset deletes. They are not
slow tests. "counts from the due date the same call sets" takes 26 ms
on a Mac, 2.5 s on a passing Windows run and over 8 s on the failing
one. One update_task call makes four fsyncs, across the note, its
agent-edit record and its creation record, and every test vault comes from os.tmpdir(), which on the runner is
C:\Users\runneradmin\AppData\Local\Temp, on the OS disk. The checkout
and RUNNER_TEMP are on D:.

A Windows job now sets TEMP and TMP to RUNNER_TEMP before anything
runs. Raising the timeout would have hidden the cost rather than
removed it. This is the likely cause, not yet a measured one: the first
Windows run on this commit is the comparison (the desktop suite took
140 s to 174 s on C:).
…n date on Windows

Delete a note permanently and make a new one under the same name within
about 15 seconds, and on Windows the new note showed the deleted note's
creation date. NTFS hands a file created or renamed into a recently
freed name the old file's birthtime (file system tunneling), and a new
note had no creation record: createNote, a first save and the MCP
server cleared any old record and left the date to the file's
birthtime. Worse, the note's first save then wrote that inherited date
into its record for good. Windows CI caught it as "removes creation
metadata on permanent deletion so a new note cannot inherit the old
date" reading a date 1.4 s old.

Every note the app makes now records now as its creation date before
the file is written: createNote, createExcalidraw, converting an
Obsidian drawing, importing an external note, duplicating (desktop and
MCP), MCP create_note and the first save of a new path. The record
replaces whatever an earlier note of that name left, is written the way
the existing records are (temp file, fsync, rename), and is skipped for
a temporary folder session, which never gets a .zennotes folder.
Deleting still removes the record.

A note made outside the app (Explorer, git) still falls back to the
file's birthtime; there is nothing to record for it. Notes that already
picked up a wrong date keep it. The new tests make stat answer with the
deleted note's birthtime for the recreated path, so every platform runs
the Windows case; all three fail on the old code.
The one change a user can see in 2.90.0 is that a note deleted and made again under the same name on Windows shows its own date (bf4f86c), with the same fix in the bundled zn 0.18.1. The CI and test commits of this release have nothing to announce.
zn 0.18.1 (ZenNotes/tui v0.18.1, ZenNotes/tui#55) records a new note's creation date the moment it makes the note, so on Windows a note made again under a deleted note's name keeps its own date, the same fix this release makes in the app and its MCP server (bf4f86c). The pin is that release's signed terminal-release.json, verified against zn-release-1 and copied byte for byte; the staged binary reports zn v0.18.1. The in-app help's CLI card names 0.18.1 and desktop 2.90.0.
…visory

npm audit listed seven low findings, all GHSA-238p-pmpm-9mq7: in KaTeX before 0.18.2, prototype pollution that already exists in the page can switch on options the author left off, such as trust. Every copy of KaTeX the app loads was 0.16.47: our own dependency in app-core, desktop, web, quicklook and the share viewer, and the ones rehype-katex, micromark-extension-math and mermaid bring. Those three still ask for ^0.16.0 in their newest releases, so a root override moves them to the same 0.19.1, and the lockfile resolves one copy for the whole tree. The production audit now reports nothing at any level.

KaTeX 0.18 prefixed its internal class names (`.tag` is now `.katex-tag`, `.base` is `.katex-base`). The app's stylesheet targets only `.katex` and `.katex-display`, which keep their names, and it passes no `strict` option, so the 0.19 strict change reports nothing new. The two equation-numbering tests and the editor smoke suite read the number by the old class and now read `.katex-tag`. Driven in the built app on a note with inline, struck, numbered, matrix, aligned and underbrace math: every formula renders in Edit and Preview, numbers read (1) and (2), no KaTeX errors, no console errors or warnings.

A user theme that styled KaTeX's internal classes would need the new names; the app ships none that do.
On the 2.90.0 gate's second full run, 'opens no preview under a finger, and still does under a pointer' found no preview within vi.waitFor's default 1 s. The preview appears in about 190 ms on an idle machine and 270 ms with every core busy, and the test passed five times alone and in two more full app-core runs; the failing run shared the CPU with the desktop suite in one turbo run. The wait now allows 5 s, so a slow runner still proves the preview opens under a pointer. The touch half is unchanged: it advances fake timers and expects no preview.
@adibhanna
adibhanna merged commit b189554 into main Nov 20, 2026
12 checks passed
@adibhanna
adibhanna deleted the v2.90.0 branch November 20, 2026 21:54
@adibhanna
adibhanna restored the v2.90.0 branch November 20, 2026 21:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant